A two-step joint model based on deep learning realizes intelligent recognition of exfoliated cells in serous effusion.

Cytological examination of serous effusion is critical for diagnosing malignancies, yet it heavily relies on subjective interpretation by pathologists, leading to inconsistent accuracy and misdiagnosis, especially in regions with limited medical resources. To address this challenge, we propose a two-step deep learning framework to standardize and enhance the diagnostic process. First, we improved the YOLOv8 model by integrating the Online Convolutional Reparameterization (OREPA) module, achieving a 93.09 % sensitivity for detecting abnormal cells. Second, we employed the Dual Attention Vision Transformer (DaViT) to classify normal cells (lymphocytes, mesothelial cells, histiocytes, neutrophils) with 98.74 % accuracy. By jointly deploying these models, our approach reduces missed diagnoses and provides granular insights into cell composition, offering a robust tool for rapid and objective cytopathological diagnosis. This work bridges the gap between AI-driven automation and clinical needs, particularly in resource-constrained settings.